Explainable Deep Learning Framework for Early Autism Detection Using Behavioral Data

Abstract

The early detection of autism spectrum disorder (ASD) is important to initiate interventions as early as possible, and the common methods of identifying the disorder are prone to rely on subjective judgments and require substantial resources of time and efforts. In the given research, we propose a configuration, where machine and deep learning methods are applied to identify ASD in toddlers and young children through behavioral survey data, as well as make the process more transparent. On two publicly accessible datasets, consisting of one on toddlers, Kaggle, and another on children, UCI, we have used a simple logistic regression, a random forest model, and a deep neural network (or multilayer perceptron, or MLP) to analyze them. Those models provided the best results, with 100 percent accuracy with the toddler population and between 98 and 100 percent with the children, demonstrating that even short survey responses could objectively identify cases of ASD. In order to be even more specific, we used SHAP (SHapley Additive exPlanations) to quantify the impact of each factor. The SHAP disaggregation identified such factors as social awareness and communication skills as the most influential in the models. Altogether, the combination of traditional machine learning with deep learning and explainable AI will result in an accurate and comprehensible system to detect ASD at a young age. This approach would not only provide doctors and parents with rapid predictions but also with clear explanations why they were obtained, which would make them build trust in using AI to check ASD. The work identifies a sound and flexible method of detecting autism earlier by experimenting on two data sets and digging into what characteristics are of most interest.